Explanatory variables

There are many factors which we think might affect the probability of conflict in a given place:

  • A history of conflict
  • Political institutions
  • Timing of elections
  • Economic development 
  • Natural resources
  • Demography
  • Geographic proximity to conflict

In order to predict conflict we try to understand how these factors might develop in the future and what impact this development will have on the likelihood of conflict. 

Many of these come from established theory in our field but unconventional predictors are included if proven useful. Choosing which of these explanatory variables to consider in order to best predict future conflict is one of the big questions for ViEWS.

Modelling techniques

We use a wide range of statistical modelling techniques.

Much of our work is based on "classical" regression analysis. It is the foundation of our (dynamic simulation approach) and has the advantages of being straightforward, easy to interpret and having well known parameter distributions. Logistic regression, a type of regression where the outcome of interest is dichotomous, is one of our main tools for predicting conflict. Linear regression is used to model non-binary phenomena such as economic development.

Much has happened in recent years with the increasing availability and ease of use of machine learning techniques. ViEWS employs these newer models in our one-step-ahead forecasting framework. It works by training models that act akin to a temporal link function, they attempt answer the question 

Given the state of the our explanatory factors today, what is the probability of conflict T months into the future?

Many types of models can be trained to answer this question. These include

  • Random forests
  • K Nearest Neighbours
  • Support Vector Machines
  • Neural Networks

Selecting, tuning and evaluating these estimators is a continuous process for ViEWS.

By combining these two approaches of classical regression in dynamic simulation with machine learning type models in one step ahead forecasting through (model ensembles) we hope to combine the best of both worlds.